Contrasting seasonality in optical-biogeochemical properties of the Baltic Sea

Optical-biogeochemical relationships of particulate and dissolved organic matter are presented in support of remote sensing of the Baltic Sea pelagic. This system exhibits strong seasonality in phytoplankton community composition and wide gradients of chromophoric dissolved organic matter (CDOM), properties which are poorly handled by existing remote sensing algorithms. Absorption and scattering properties of particulate matter reflected the seasonality in biological (phytoplankton succession) and physical (thermal stratification) processes. Inherent optical properties showed much wider variability when normalized to the chlorophyll-a concentration compared to normalization to either total suspended matter dry weight or particulate organic carbon. The particle population had the largest optical variability in summer and was dominated by organic matter in both seasons. The geographic variability of CDOM and relationships with dissolved organic carbon (DOC) are also presented. CDOM dominated light absorption at blue wavelengths, contributing 81% (median) of the absorption by all water constituents at 400 nm and 63% at 442 nm. Consequentially, 90% of water-leaving radiance at 412 nm originated from a layer (z90) no deeper than approximately 1.0 m. With water increasingly attenuating light at longer wavelengths, a green peak in light penetration and reflectance is always present in these waters, with z90 up to 3.0–3.5 m depth, whereas z90 only exceeds 5 m at biomass < 5 mg Chla m-3. High absorption combined with a weakly scattering particle population (despite median phytoplankton biomass of 14.1 and 4.3 mg Chla m-3 in spring and summer samples, respectively), characterize this sea as a dark water body for which dedicated or exceptionally robust remote sensing techniques are required. Seasonal and regional optical-biogeochemical models, data distributions, and an extensive set of simulated remote-sensing reflectance spectra for testing of remote sensing algorithms are provided as supplementary data.


Introduction
The Baltic Sea is a brackish, semi-enclosed, and relatively shallow (average depth approximately 55 m) coastal sea with large latitudinal gradients in salinity and dissolved organic matter (DOM). Seasonality of the physics and hydrodynamics of the sea are driven by a  β p (λ) Particulate backcattering-to-scattering ratio at waveband λ (spectral average when λ is omitted) dimensionless S Spectral slope coefficient of a CDOM (λ) model (Eq 1) nm -1 or μm -1 (see text)

SIOP Specific inherent optical properties
a Ã X ;Chla ðlÞ Chla-specific a x (λ) m 2 mg -1 a Ã X ;TSM ðlÞ TSM-specific a x (λ) m 2 g -1 a Ã X ;POC ðlÞ POC-specific a x (λ) water layer is markedly shallower than, but under mixed conditions still representative of, the mixed layer. Under calm wind conditions, motile or buoyant phytoplankton may overcome vertical mixing and form relatively dense layers, including cyanobacterial surface accumulations for which the Baltic Sea is notorious [18]. Optical modelling previously demonstrated that the vertical distribution of phytoplankton biomass in the upper water layer does indeed significantly impact water reflectance [23]. During thermally stratified periods (in summer), detrital matter may additionally linger in the observed water layer. Lacking in situ observations representative of various degrees of vertical mixing, it is unknown to what extent remote sensing and (optical) models can accurately represent these conditions. Studies of the optical properties of the system should therefore extend to both inherent and apparent optical properties (IOPs, AOPs), and inspect whether phytoplankton is distributed homogeneously above and below the limit of remote observations. Given the poor environmental status of the Baltic Sea ecosystem, efforts into operational optical monitoring and analysis of past and future satellite data are needed to support its costeffective management. To this end, remote sensing algorithms that are optimally suited to the unique optical conditions of the Baltic Sea are needed. In a broader perspective, tackling the challenges in remote sensing of the dark, yet productive Baltic Sea waters will support increasingly robust algorithm development for optically extreme waters elsewhere. The new generation of operational remote sensors in the Copernicus Sentinel mission will facilitate new algorithm strategies, notably with the introduction of blue (400 nm) and red (674 nm) wavebands on the Ocean Land Colour Instrument (OLCI) on the Sentinel-3 mission, which should prove useful to characterize and quantify CDOM and phytoplankton Chla, respectively. A better understanding of the spatiotemporal variability of optically active substances is the first step in this direction and the primary focus of the present work. Because detailed optical in situ measurements from the open sea are scarce, a reference database of optical properties and associated water colour is needed for algorithm testing and development. Detailed and seasonally specific (where required) models, distributions of the underlying biogeochemical parameters, and a set of simulated remote-sensing reflectance (R rs ) are therefore provided with this paper.

Sampling strategy
Optical and biogeochemical observations were compiled from eight research cruises in spring and summer in the period 2008-2012. 145 station visits were made with RV Aranda resulting in 459 surface water (0-17 m) samples taken with Niskin bottles on a sampling rosette equipped with a calibrated conductivity, temperature, and depth (CTD) instrument. When wave conditions allowed, samples were taken from just under the surface (0.5 m), subsurface (3 m), and depths corresponding to the top (nominally 10 m) and bottom (nominally 15 m) of any thermocline present. Under rough weather the surface samples could not always be collected. When streaks of buoyant phytoplankton (cyanobacteria) were present the research vessel was let to drift to minimize disturbance of shallow stratification. Surface accumulations were occasionally encountered in summer but always took the form of easily disturbed films rather than thick layers. Wind and wave conditions or busy ship traffic frequently necessitated the use of thrusters to stay the vessel, particularly during deep profiling. Water samples and measurements obtained close to the surface may therefore underrepresent the extent of vertical stratification of the water column.
In addition to depth profiling stations, 72 water samples were obtained while cruising from a dedicated flow-through system on board the RV with a water intake at 3 m depth. Additionally, 75 water samples collected with an automated refrigerated (4˚C) water sampler on ship-ofopportunity MS Finnmaid of the Alg@line network were included from 17 transects between Travemünde and Helsinki, using up to 12 water samples per transect. The latter samples were primarily used to characterize the absorption by CDOM.
A map of sampling locations is given in Fig 1. Samples from less-frequently visited were aggregated with adjacent areas as indicated in Table 2. The Northern Baltic Proper and the western half of the Gulf of Finland are best represented in the data set, with visits in every field campaign. For the southern Baltic Proper and the Gulf of Bothnia (Bothnian Sea, Quark, and Bothnian Bay) all samples were taken in summer with the exception of some transects of MS Finnmaid. Russian waters in the Eastern Gulf of Finland were only visited in August 2009. No data were collected in the Gulf of Riga and the Danish straits and only two stations are located in the Western Gotland Basin sampled in August of 2008 and 2010. Throughout this paper when referring to sample sets as 'spring' these were collected during cruises in April, 'summer' samples included July as well as August, unless specifically marked as mid-summer (July) or late summer (August). Samples taken in other months were only used to characterize CDOM.
Sampling permits were not required for operations of the Finnish research vessel in the territorial waters of Finland. In all other territorial seas and exclusive economic zones sampling permits were obtained prior to each research cruise from the respective national authorities: the Swedish coast guard, the Estonian Ministry of Foreign Affairs, the Federal Agency for Science and Innovations and/or the Ministry of Foreign Affairs of Russia, the Polish Ministry of Foreign Affairs, and the Federal Maritime and Hydrographic Agency of Germany.

In-water measurements
Multiple depth profiles were taken at each station with a set of active optical sensors including a Wetlabs AC-S spectral absorption and beam attenuation meter with 0.1-m path length, a  Table 2 Wetlabs volume scattering meter (VSF, 470, 532, 660 nm each at angles 100˚, 125˚, 150˚) and a Wetlabs BB3 fixed-angle backscattering meter (412, 595, 715 nm at 117˚). A second BB3 (470, 532, 715 nm) was used in some cruises, in which case the results of overlapping wavebands were averaged. The deployment depth of each sensor was determined from its vertical distance to a pressure sensor on an attached CTD instrument (RBR XR-620 or Sea-Bird SBE 37-SI MicroCAT). AC-S measurements were processed to yield spectral absorption, beam attenuation, and scattering coefficients (respectively a(λ), c(λ), and b(λ) as c(λ)-a(λ)) by subtracting measurement blanks of highly purified water (measured before and after every deployment), and applying temperature and salinity corrections for the absorption and scattering of seawater. AC-S profiles were additionally subjected to visual inspection, removing measurements affected by bubbles or sensor movement causing timing errors of the internal filter wheel. Several depth profiles in early spring had to be removed due to condensation on internal optics, caused by very low (< 0˚C) water temperatures. VSF and BB3 measurements were corrected for absorption ('sigma correction' following the manufacturer's recommendations) and subsequently converted from fixed-angle scattering measurements to b b (λ) at six wavelengths (sensors combined) following integration of a polynomial function over the angle-specific measurements of the VSF, and further (and for the BB3) following [24]. Depth profiles were collected over at least the first 15 m and up to 45 m depth, simultaneously with water sampling from the sampling rosette whenever weather conditions allowed this, or otherwise within 0.5 h. Absorption measured from AC-S was used to correct other in-water measurements, whereas AC-S scattering measurements were integrated over depth bins of 1 m for comparison against absorption measured from discrete water samples (described below).
At daylight stations, spectral up-and downwelling irradiance E u (z,λ) and E d (z,λ) were recorded with TriOS Ramses-ACC-VIS sensors (320-950 nm) as a function of depth, using a reference sensor on deck to account for fluctuations in the solar irradiance arriving at the water surface. Irradiance measurements were taken from 0-15 m depth with the ship positioned so that sensors were exposed to direct sunlight. Sampling depth was measured with a pressure sensor (SBE-50, Sea Bird Electronics) attached to the sensor cage. A self-shading correction using a(λ) obtained with the AC-S was applied to measurements of E u (z, λ). The downwelling diffuse attenuation coefficient K d (λ) was calculated as the exponential slope fitted through depth profiles of E d (z,λ). Irradiance measurements taken over 0-7 m were used to determine attenuation of the photon flux density of photosynthetically active radiation (PAR) by first determining PAR at depth and subsequently fitting K d (PAR). Extrapolation of E u (z,λ) and E d (z,λ) profiles to zero-depth allowed calculation of the dimensionless subsurface irradiance reflectance. Secchi disk depth (Z SD , m) was recorded with 0.5-m precision at every sampling station by lowering the 30-cm white disk from the research deck (approximately 4 m above the water line), avoiding sun glint and ship shadows.

Laboratory analyses
Water samples were processed within two hours from sampling and kept at light intensity and temperature similar to their sampling origin until further laboratory analysis. Particulate material was concentrated onto duplicate Whatman GF/F glass fibre filters with a nominal pore size of 0.7 μm. Filters for gravimetric analysis of the dry weight of total suspended matter (TSM) were first rinsed with ultrapure water, combusted four hours at 450˚C, and individually weighed prior to each cruise. Salt crystals were removed from the filters by filtering at least 50 ml of ultrapure water through the filter following filtration of sea water and before removing the filters from the filtration manifold and storing them dry in vented petri dishes. Filters for particulate organic C, N and P were acid-rinsed before washing and combustion. Sample filtrates were collected using disposable membrane filter cartridges (pore size 0.2 μm) and polycarbonate syringes without silicone stoppers.
Absorption of particulate matter was determined using the quantitative filter pad technique [25,26] using Whatman GF/F glass fibre filters (0.7 μm nominal path length). The filters were placed at the centre of the 150-mm integrating sphere accessory of a PerkinElmer Lambda800 spectrophotometer for absorbance scans at 1-nm intervals, 2-nm slit width, between 300 and 800 nm (350-800 nm for the 2008 summer cruise). Scans were converted to the absorption coefficient of particulate matter (a p (λ), units m -1 ) following a previous characterization of path length amplification for glass fibre filters at the centre of this integrating sphere, using Baltic Sea samples [27]. Further separation of a p (λ) was achieved by bleaching the filters, isolating phytoplankton pigment absorption (a ph (λ)). The bleaching procedure consisted of an 8-min exposure to a 0.1% solution of sodium hypochlorite [28,29], if necessary followed by a 70˚C treatment of 80% ethanol for 5 min. The remaining fraction commonly referred to as 'nonalgal particulates' (a nap (λ)) includes living and detrital algae and cyanobacteria, as well as other biotic particles and mineral particulate matter.
The a CDOM (λ) spectrum was measured by passing samples through 0.2-μm membrane filters and subsequent spectrophotometric determination in quartz cuvettes, against a reference of ultrapure water. Spectra were recorded over the 190-800 nm range and converted to units of absorption (log-10 to natural log conversion and normalization to 1-m path length). Absorbance results from 1 cm and longer (4, 5, or 10 cm) cells were combined to yield the best signal / noise relationship in both the ultraviolet (UV, high CDOM absorption: shorter cells) and visible to near infra-red (NIR) range.
Chla was extracted in 10 mL 96% ethanol for 24 h in darkness at room temperature [30,31]. Chla concentration was quantified on a Cary Varian Eclipse spectrofluorometer calibrated with pure Chla (Sigma) and excitation and emission slits of 5 nm centred at 430 and 670 nm, respectively. Filters for particulate organic Carbon (POC), nitrogen (PON), and phosphorus (POP) were allowed to dry and stored at room temperature (20˚C) until analysis. POC and PON were measured from the same filter with a mass spectrometer (Europa Scientific). POP was determined according to [32]. Filtrates for DOC were measured by high-temperature catalytic oxidation and non-dispersive infra-red detection using a total organic Carbon analyzer (Shimadzu TOC-VCPH) equipped with chemiluminescence detector (Shimadzu TNM-1). Filters for TSM dry weight were dried overnight at 60˚C and subsequently weighed before and after combustion of organic material (4 hours at 450˚C) to yield total and the inorganic fraction of dry weight.
In addition to sampling stations, underway samples were taken from a dedicated flowthrough system sampling at approximately 3 m depth. Chla was always quantified from these samples, whereas CDOM absorption, HPLC pigments, particulate dry weight, and organic fractions of dissolved C and N and particulate C, N and P were included depending on the specific purpose of each cruise.
Phytoplankton >2 μm were counted from 29 water samples taken during two spring cruises in 2011-2012 and 70 water samples taken during summer cruises in 2010-2012. Samples were taken from Niskin bottles at 3 m depth, fixed in acidic Lugol's medium, stored in the dark at 4˚C and left to settle in an Utermöhl chamber before examination by inverted light-microscopy. Cell numbers were converted to biomass using conversion factors of the HELCOM Phytoplankton Expert Group [33].
Phytoplankton biomass in spring significantly exceeded summer biomass when expressed in terms of Chla (Student's t-test assuming unequal variance, p < 0.001), TSM (p = 0.026), or POC (p < 0.001). Shapiro-Wilk tests on log-transformed data confirmed normal distributions (p > 0.01) for summer Chla, spring TSM and ITSM/TSM, spring and summer POC, and summer and all-season PON:POP ratios. Histograms of these properties allow further comparison of seasonal differences (Fig 3). Basic statistics (range, mean, median, quartiles) on these distributions are provided in the supplementary S1 Table. Spring particle communities were associated with sharply peaked ISM/TSM ( Fig 3C) and significantly lower median PON:POP ratios (p < 0.001, Fig 3E) compared to broader ranges in summer, indicative of spring excess and summer depletion of bioavailable phosphorus.

Biogeochemical relationships
Autochthonous production of organic matter dominated the particle population during productive periods in the open waters of the eutrophic Baltic Sea, where turbidity caused by mineral particles was already low (Fig 3B and 3C). Significant relationships can therefore be observed between POC and Chla in both spring (r 2 = 0.41, p < 0.001, n = 75) and summer   (r 2 = 0.64, p < 0.001, n = 134) as well as in samples from both seasons combined (r 2 = 0.51, p < 0.001, n = 209). The average POC:Chla ratio in summer was nearly twice as higher as in spring (Fig 4A). Relationships between TSM and Chla were weak in comparison (Fig 4B), significant but weak in summer samples (r 2 = 0.11, p < 0.001, n = 174) but lacking a meaningful correlation in spring. Detailed analysis of differences in the POC:Chla ratio observed between the seasons (Fig 4A) revealed a significant correlation with the molar PON:POP ratio, with a relatively continuous distribution between the seasons (Fig 4C), compared to the seasonally clustered relationships between Chla and either POC or TSM.

Diffuse attenuation relationships with reflectance and Secchi depth
In the visible spectrum, variability in K d (λ) was lowest between 600-700 nm (20%) and highest between 400-500 nm (38%). Phytoplankton pigment visibly influenced K d (675) in spring samples with high phytoplankton biomass (Fig 5A), whereas increase in K d (λ) at 600 nm and 700 nm corresponds to a rise in a w (λ). Despite considerable variability in the amplitude of K d (λ) its spectral shape was highly conserved between samples, irrespective of season. K d (λ) normalized to K d (PAR) had 10% variability at 600-700 nm and 17% at 400-500 nm ( Fig 5B). The mean K d (λ):K d (PAR) ratio demonstrates that K d (501) = K d (PAR) within 1%, whereas K d (490), the waveband at which K d is most commonly reported in remote sensing applications, exceeded K d (PAR) on average by 10%.
Subsurface irradiance reflectance spectra are plotted in Fig 5C. The highly conserved shape of K d (λ), with the absorption by CDOM and water consistently contributing to high attenuation in the short and long wavelength ends of the spectrum, resulted in reflectance spectra which were consistently low in these areas, with peak reflectance in the green spectral domain.
Models to predict K d (PAR) from more commonly observed Secchi disk depth were previously evaluated to hold the form K d (PAR) = f / Z SD with f empirically found close to 1.7 ( [36] and references therein). In the current data set Z SD varied 2-12 m (n = 38) in spring and 2-7.5 m (n = 101) in summer. Fitting the model with f = 1.7 to the data presented here resulted in r 2 = 0.48 and RMSE = 0.14 m, which is close to the optimal fit obtained with f = 2.14 (r 2 = 0.48 and RMSE = 0.11 m, plotted in Fig 5C). When only spring samples were considered f = 2.44 improved the correlation but not the average error (r 2 = 0.74, RMSE = 0.11) while for summer Optical seasonality of the Baltic Sea samples there was no marked difference using the best fit of f = 1.98 against the previously published model with f = 1.7 (r 2 = 0.45, RMSE = 0.09). It is noted that these errors are small with respect to the precision of 0.5 m with which Z SD was observed at sea. Alternative models (e.g. slope and intercept, or exponential fitting) were able to improve the fit towards clearer waters but without improved error properties over the whole range (not shown). Salinity is a poor proxy for model coefficient f (Fig 5D): a weak negative correlation with large associated error was observed between f and salinity in summer (r 2 = 0.34, RMSE = 3.8), while the trend was opposite for spring (r 2 = 0.30, RMSE = 3.4).  Table. https://doi.org/10.1371/journal.pone.0173357.g005

Vertical distributions of Chla and POC and remote sensing perception
Vertical profiles of Chla analysed from bottle samples (Fig 6) illustrate the degree of vertical inhomogeneity encountered in the water column. It is noted that, despite precautions taken to minimally manoeuvre the ship during sampling, ship-induced mixing may at times dampen vertical biomass profiles. Inhomogeneous biomass distribution was observed at some spring stations ( Fig 6A) and more frequently in mid and late summer (Fig 6B and 6C). The ratio POC:Chla was relatively constant over the first 10 m depth ( Fig 6D) which is consistent with results shown in Fig 4A. Light penetration depth is overlaid in Fig 6A-6C to illustrate how representative a remotely sensed signal is of column biomass. The first optical depth (z 90 ) is the inverse of the vertical diffuse attenuation coefficient K d , i.e. the layer depth from which 90% of the water-leaving radiance observed by a remote sensor originates [37]. The Baltic Sea attenuates blue (shown at 412 nm) and red (674 nm) light (Fig 5A and 5B) with the highest efficiency. Green light (shown at 560 nm) penetrates deepest and has the strongest dependence on phytoplankton biomass. Mean z 90 values were, irrespective of season, in the order of 1.0 m (blue), 1.5 m (red), and 3.0-3.5 m (green) and only exceeded 5 m in the green band when biomass was < 5 mg Chla m -3 . Non-parametric testing of variance in z 90 between spring, mid-and late summer only showed significant differences (irrespective of wavelength) between spring and mid-summer (Kruskal-Wallis test, F = 1.007, p = 0.32). If photosynthesis is considered feasible up to the 1% light penetration depth (z eu = 4.6 × z 90 ), this yields a representative maximum light penetration for photosynthesis in the order of 13.8-16.1 m.
Assuming the signal above z 90 to be representative of the full euphotic layer leads to underestimates of photosynthetic activity when biomass deeper than z 90 exceeds biomass above this depth, and reciprocally (and more commonly) peak biomass above z 90 may cause overestimation of column productivity if remotely sensed biomass estimates are used while the extent of mixing is not known. Table 3 lists the occurrence of peak biomass in the top (< 3 m), In 44% of spring stations Chla biomass was equally (< 10% variation) distributed in the water column, compared to 16-25% of stations visited in mid and late summer. In 39% of spring stations and 25% mid-and 35% late-summer stations a peak in Chla biomass was found near the surface. In terms of POC, deeper (> 10 m) peaks of organic matter were found more frequently, particularly in mid-summer (18%).

Light absorption budgets
When excluding the influence of a w (λ), the absorption budget of the Baltic Sea pelagic observed during spring and summer seasons was dominated by a CDOM in blue (λ < 442 nm) and by a ϕ at red wavebands (λ = 665-681 nm), as visualized with ternary plots for visible and NIR bands of Sentinel-3 OLCI (Fig 7). A gradual shift from a CDOM to a ϕ dominance occured between blue and red bands, where a NAP occasionally contributed up to 50% in the green-yellow domain (560-620 nm). Median a NAP ranged 5-12% (both seasons) between 400-700 nm. Median a CDOM was 76% (±8%) at 412 nm (irrespective of season), i.e. between the > 90% reported in [5] in the Gulf of Bothnia in August (absent phytoplankton bloom), and the 38-70% range reported for the area influenced by the Oder river in autumn [6]. Median a CDOM amounted to 81% (±7%) at 400 nm and 63% (±10%) at 442 nm. The relative contribution of a ϕ was 84-87% in bands 665, 674, and 681 nm in spring but only 68%, 74%, and 76% respectively in these bands during summer. The difference between the seasons was not due to stronger relative contribution of a NAP (median in the order of 7% in both seasons) but due to stronger a CDOM contribution with a median around 16% (±16%) in summer compared to 5% (±12%) in spring samples.

Chromophoric dissolved organic matter absorption relationships
The shape and magnitude of a CDOM (λ) determined from 264 samples (251 stations, April-December, Fig 8A) could be adequately described using exponential slope coefficient S and a CDOM at reference wavelength λ 0 , following the commonly used exponential model (see e.g. [38]), allowing for any residual noise or scattering with intercept k [39]: The mean a CDOM (λ) spectrum was well captured by the model over the 400-700 nm range using λ 0 = 412 nm (r 2 = 1.00), with S(400-700) = 0.0177 nm -1 and k = 0.030. Considerable variability in the UV region reflects the variable presence of chromophores from terrestrial and autochthonous origins, with associated peaks in the standard deviation spectrum (Fig 8A). This variability is not observed in the visible domain (Fig 8A inset) and therefore not explored here. S is here expressed as the slope measured over the 400-700 nm range. Variability in S ( Fig 8B) and a CDOM (412) (Fig 8C) was substantial, showing a short-tailed distribution of S (mean = 0.0177, median 0.0182 nm -1 ) and a bimodal distribution of a CDOM (412), the latter having modes at 0.53 m -1 (central and southern sea areas) and 0.85 m -1 (Gulf of Bothnia, Eastern Gulf of Finland).
Relationships between a CDOM magnitude and S are expected to have a geographical component due to variant run-off from river catchments and subsequent dilution and transformation in the lifetime of CDOM in the sea. In the Baltic Sea the CDOM riverine input is particularly strong in the eastern Gulf of Finland and the semi-isolated Gulf of Bothnia. Indeed, three models were found to represent the relation between a CDOM (412) and S(400-700) between these   Table. https://doi.org/10.1371/journal.pone.0173357.g008 (r 2 = 0.52, p < 0.001, RMSE = 0.21 m -1 , n = 33), although the low number of samples and some discontinuity in a CDOM (412) warrants caution in using this model. Interestingly, samples from the eastern Gulf of Finland (in the Neva river bay) with a CDOM (412) > 2.5 m -1 are better described by this model than by the linear model for all samples taken in the Gulf of Finland (Fig 9A, green line), which has the shape a CDOM (412) = -137.66 S + 3.60. The GoF model is the weakest of the three models (r 2 = 0.22, p < 0.001, RMSE = 0.40 m -1 , n = 84) and illustrates the presence of complex mixing dynamics with multiple riverine inputs to this area [11].
Salinity and a CDOM (λ) can be used interdependently to trace water masses in coastal seas (e.g. [40]). Correlations between a CDOM (λ) and salinity have been established in the southern Baltic Sea [12,28] and to trace water masses in the Baltic-North Sea exchange [41,42]. In this census of optical properties of the open Baltic Sea, a CDOM (λ) and salinity are fairly well correlated (Fig 9B, r 2 = 0.63, p < 0.001, RMSE = 0.25 m -1 , n = 240) although a high degree of scatter is observed in the western Gulf of Finland. Some degree of seasonality, as previously found for waters of the southern Baltic Sea [12], will be evident but remains to be studied in detail, including variability in absorption properties in the UV. Exploration of correlations between S and salinity or between a CDOM (412), S, and the distance to major rivers did not yield patterns which could be exploited in statistical optical models.
The ultimate applications of a CDOM (λ) mapping coupled to biogeochemical ecosystem modelling are to predict underwater light availability for photosynthesis, and to trace DOC. Once a CDOM is retrieved from a remote optical sensor at a reference (blue) waveband, the models described above for S and salinity will be useful to determine a CDOM (λ) with relatively high confidence. Subsequently, a model relating DOC to a CDOM is required. Two models could be described to capture the variability in this relationship.

Particulate absorption relationships and pigment packaging
Seasonal differences in particulate SIOPs were most prominent for absorption expressed per unit of Chla, i.e. as a Ã f; Chla ðlÞ ( Fig 10A and 10B) and a Ã nap; Chla ðlÞ ( Fig 11A and 11B). Mass-normalized results for a p (λ), the sum of a ϕ and a nap , are not plotted here but supplementary S7, S8 and S9 Tables include SIOP data for all of a Ã p ðlÞ, a Ã f ðlÞ, and a Ã nap ðlÞ. Spring and summer a Ã f;Chla ðlÞ were significantly different (t-test, unequal variance, two-tailed p-value << 0.001) throughout the spectrum.
Comparing mean a Ã f;Chla ð440Þ of spring and summer particle populations, we observed a Ã f;Chla ð440Þ twice as high in summer (0.040 (±0.010) m 2 mg -1 ) compared to spring (0.020 (±0.006) m 2 mg -1 ). The difference was smaller at 675 nm with 0.014 (±0.003) m 2 mg -1 in spring compared to 0.021 (±0.004) m 2 mg -1 in summer. The variation in the spectral magnitude (average a Ã f;Chla over 400-700 nm) was similar between the seasons with a coefficient of variation (CV) of 26% in spring (n = 133 samples from 36 stations) compared to 28% (n = 302 samples, 94 stations) in summer. The presence of phycoerythrin (absorption peak at 560 nm) was noted only in summer, likely associated with Synechococcus sp. This pigment peak was also more distinct in samples taken at the bottom of the euphotic zone (10-20 m), although depth was otherwise a minor source of variability in absorption properties and it is not explored further here.
A wider range of variability was observed in a Ã f; TSM ðlÞ ( Fig 10C and 10D) with spectrally averaged CV = 50% in spring and 392% in summer. There was no statistical difference (t-test) at any wavelength between the spring and summer sets and mean a Ã f; TSM ð440Þ in spring (0.100 m 2 g -1 ) and summer (0.099 m 2 g -1 ) were near identical, while summer a Ã f; TSM ð675Þ was lower (0.052 m 2 g -1 ) than in spring (0.072 m 2 g -1 ) which contrast the results given for a Ã f; Chla ð675Þ. Normalization to POC (Fig 10E and 10F) yielded even smaller differences between mean spring and summer spectra with a Ã f;POC ð440Þ ¼ 0:006 m 2 mmol -1 in spring and 0.007 m 2 mmol -1 in summer, and a Ã f;POC ð675Þ ¼ 0:004 m 2 mmol -1 in both spring and summer. A large degree of variability (CV = 84%) in the magnitude of a Ã f;POC was observed in summer (CV = 27% in spring), which may cause the difference between spring and summer a Ã f;POC ðlÞ sets to be statistically significant (t-test, unequal variance, two-tailed p < 0.001) at all wavebands except for green (550-580 nm) and NIR minima, despite the similarity in shape and magnitude of the seasonal mean spectra.
Variability in a Ã nap (Fig 11) was higher in summer compared to spring, regardless of whether normalization followed the concentration of Chla (summer CV = 195%, spring 59%), TSM (summer CV = 785%, spring 57%) or POC (summer CV = 158%, spring 69%). The spectral  Table. https://doi.org/10.1371/journal.pone.0173357.g010 Optical seasonality of the Baltic Sea variability of a nap (λ) was modest in both seasons. The spectral slope of a nap (λ), fitted over 400-700 nm using the same model as for a CDOM (λ) (Eq 1), varied in the range 0.0049-0.0126 nm -1 in spring with median 0.0097 nm -1 and mean (± standard deviation) of 0.0089 (± 0.0016) nm -1 . In summer, the range spanned 0.0031-0.0144 nm -1 with a median value of 0.0103 nm -1 and mean 0.0098 (± 0.0021) nm -1 . Summer samples did occasionally show a less uniform decline of absorption with wavelength in the blue domain, which may suggest remnant pigment absorption despite rigorous bleaching procedures. Detailed data distributions of the slope of a nap (λ) are given in the supplementary S2 Table. Pigment packaging is expressed as a reduction in pigment absorption efficiency with increasing cellular pigment concentration. While the influence of cell size and morphology is not explored in this study, the wide range of pigment concentrations encountered in the data set makes it possible to express a spectral model for the effects of pigment packaging on a Ã p;chla ðlÞ and a Ã f;chla ðlÞ. The effect is strongest when a Ã p;chla ðlÞ is considered, as shown in Fig  12A and 12B at 440 and 675 nm, respectively. Spring and summer response functions can be seen to deviate at low and intermediate Chla concentrations (0.5-5 mg m -3 ). Power function regression fits capture the response at both plotted wavebands with acceptable coefficient of determination and error properties, although summer samples dominate the model that is given for the seasons combined.  Table. https://doi.org/10.1371/journal.pone.0173357.g011 Optical seasonality of the Baltic Sea The best-fitting power function applied to spring observations of a Ã f;chla ðlÞ (Fig 12C and  12D) corresponds well to the frequently applied model (B95) of [43]-either model would result in a root mean square error (RMSE) of 0.005 m 2 mg -1 for a Ã f;chla ð440Þ. The error increases approximately three-fold to 0.014 m 2 mg -1 when both spring and summer observations are considered against B95, compared to 0.010 m 2 mg -1 for the best fitting model. The B95 model did not significantly describe the package effect in summer observations alone at 440 nm. The model errors are smaller when a Ã f;chla ð675Þ is considered (Fig 12D). The B95 model then has an associated error of 0.005 m 2 mg -1 against all samples combined, compared to 0.004 m 2 mg -1 of the best-fitting model. This error is likely acceptable, particularly considering the high amount f ;chla ð675Þ. Power model fits to spring, summer, and observations in both seasons combined plotted as blue, green, and black lines, respectively. Shaded areas in corresponding hues indicate 95% confidence limits for these models. All models were significant at p < 0.001. The model for a Ã f ;chla ðlÞ of [43] is included for reference (B95) in panels C and D, along with RMSE corresponding to fits of the present data (spring and all combined) to this model. Supplementary S11 Table contains regression fits at all wavelengths. https://doi.org/10.1371/journal.pone.0173357.g012 Optical seasonality of the Baltic Sea of scatter around of the data around the models and low associated coefficients of determination (r 2 = 0.14, 0.07, and 0.32 respectively for spring, summer, and both seasons).

Particulate (back)scattering relationships
Particulate scattering b p (λ) in spring showed a negligible spectral slope but a distinct influence of pigment absorption, with ±10% variation around the spectral average. Spectral variability was most visible at the blue and red peaks of Chla and readily observed in plots of b Ã p ðlÞ (Fig  13), similar to earlier reports [44][45][46]. The relative spectral variation around pigment absorption peaks was minor in summer, probably owing to higher magnitude of b Ã p ðlÞ. A more distinct spectral slope of -0.11% nm -1 (between 400 and 700 nm) was, however, observed in summer samples.
Despite clear differences in the magnitude of b Ã p ðlÞ between the seasons, within-season variability was similar. In spring, the spectrally averaged CV was 56%, 42%, and 56% for normalization of b p (λ) to Chla, TSM, and POC, respectively. In summer, b Ã p;chla ðlÞ was equally variable (56%), whereas b Ã p;tsm ð86%Þ and b Ã p;POC ð75%Þ had wider variability. Chla concentration is thus a consistent indicator of the variability in the particulate scattering SIOP in both bloom seasons. If particles with a high inorganic content played a significant role in scattering, we would  Table. https://doi.org/10.1371/journal.pone.0173357.g013 Optical seasonality of the Baltic Sea expect b Ã p ðlÞ to show an increase with inorganic particulate dry weight. Instead, a weak (negative) trend of b Ã p;TSM ð532Þ with an increasing proportion of ITSM is observed, and only when summer samples are considered (Fig 14, r 2 = 0.30, p < 0.001, RMSE = 0.16 m 2 g -1 , n = 168).
Spectral variation in particulate backscattering (b bp (λ)) is presented as boxplots (Fig 15) to accommodate for incomplete overlap in the waveband sets of VSF and BB-3 sensors deployed during various cruises. Backscattering spectra were similar in shape to b Ã p ðlÞ shown in Fig 13, Table. https://doi.org/10.1371/journal.pone.0173357.g014  Table. https://doi.org/10.1371/journal.pone.0173357.g015 with influence of pigment absorption particularly visible in the red channel at 660 nm. A minor peak at 595 nm is visible from observations in spring. The observed range is similar to that reported previously for the Baltic Sea in the Gulf of Finland [47]. Spring and summer b bp spectral means were not significantly different. The particulate backscattering ratio (β p (λ) = b bp (λ)/b p (λ)) was calculated for all wavebands where b bp (λ) was available. Spectral variation identified in b bp (λ) and b p (λ) was amplified in β(λ) as shown in Fig 16A and 16B. Higher b Ã p ðlÞ, offset against generally lower constituent concentrations in summer (Fig 3) and seasonally similar b p (λ) resulted in net higher backscattering ratios in spring than in summer, with spring β(532) close to the commonly used values of Petzold's volume scattering function for San Diego Harbor at the same angle and wavelength [48]. Both the spectral mean ( Fig 16C) and spectral slope (Fig 16D) of β were significantly different between seasons.
Weak but significant season-specific correlations were found between b bp (532) and Chla, POC, and TSM (Fig 17). Seasonal effects were least evident with TSM and POC. The poorest correlations were found with the inorganic fraction of suspended matter (ISM/TSM), suggesting an absence of inorganic particles in most observations, while mineral matter may be contributed by silicates from diatom frustules present in the ash fraction.

Discussion
The Baltic Sea is perhaps most widely known for its occasional magnificent displays of cyanobacterial blooms, decorating surface waters during periods of calm weather in summer, and clearly visible from space [18,49,50]. The prolonged calm weather leading to surface accumulation [17] is however not as common as occasional wind mixing, and summer blooms are Optical seasonality of the Baltic Sea therefore usually dispersed between the water surface and thermocline (Fig 6, Table 3). Optical observations presented in this study include relatively few encounters with cyanobacteria accumulated at the surface. Correspondingly, we have not observed marked optical variability over depth within the euphotic zone. Marginally wider variation with depth was observed in relation to the absorption signature of the phycoerythrin pigment in summer, likely associated with pico-cyanobacteria [51]. This, however, will have minimal impact on remote sensing studies observing only the first optical depth. The prevalence of both filamentous and picocyanobacteria in summer is nevertheless a prominent factor in the optical seasonality of the Baltic Sea. Optical seasonality of the Baltic Sea Spring blooms develop significantly higher POC in the mixed layer than summer blooms ( Fig 3D) and even three-fold higher Chla (Fig 3A). The difference in the Chla to POC ratio is explained by a less prominent role of Chla in photosynthetic light harvesting in cyanobacteria (which are only prevalent in summer) compared to algae [51][52][53]. Limited variability in spring a Ã f and a Ã nap SIOPs, regardless of whether normalization of the IOPs follows Chla, TSM, or POC concentration (Figs 10 and 11), is a strong indication that absorption is governed by the phytoplankton component. Variability in summer IOPs is better explained by POC and Chla than by TSM concentration, supporting our understanding that phytoplankton are the main source of variability in particulate absorption, while detrital matter is a more variable component in summer. We also observe that within-season correlations between Chla and TSM are well established (Fig 4B) and inorganic TSM does not act as a prominent driver of scattering SIOPs.
The Baltic Sea spring bloom builds up as light availability increases, and dissipates soon after nutrient depletion as particles sediment out. Consequently, the spring bloom has an intense but short-lived impact on the optical properties of the surface layer [54,55]. In contrast, thermal stratification in summer causes the summer particle population to linger in surface waters. We understand this to lead to a relatively complex and variable particle composition in summer compared to our observations during spring bloom. We observed a higher proportion of absorption associated with the non-algal particle fraction of particles (Fig 7), wider variability in IOPs between stations (Figs 10, 11 and 13) and a relatively diverse phytoplankton community (Fig 2) in summer compared to spring. Inorganic nutrient availability, particularly bio-available phosphorus, is typically at minimum detection levels during summer sampling (not shown), suggesting that after the initial establishment of the phytoplankton population its biomass is fuelled by recycling of nutrients. The more significant contribution of a nap (λ) to light absorption in summer compared to spring is therefore likely associated with a larger proportion of detrital material, in turn contributing to nutrient recycling and prolonged blooms. Reciprocally, our optical measurements support our understanding that particulate matter produced in spring is mixed deeper and met with faster export from the visible surface layer. It would be of interest to bio-optical modelling and remote sensing efforts to further differentiate the absorption and scattering SIOPs of phytoplankton, detritus, and mineral particulates so that the influence of seasonality in phytoplankton optical properties and the prevalence of detrital material can be investigated separately. This will likely require a multivariate statistical approach to an even larger and seasonally resolved data set than presented here.
Absorption of light by CDOM dominates the underwater light climate even during the most productive periods in the Baltic Sea and has a marked influence on the absorption by water constituents, even in satellite wavebands in the 600-665 nm range (Fig 7). Given that (1) the predominantly organic particle population observed here does not give rise to strong particulate backscatter, and (2) a CDOM in the blue-red and a w at >600 nm provide an efficient absorption medium and weak water-leaving radiance, reflectance is always expected to peak in the green domain between 560-600 nm. Reflectance simulated from the presented SIOP models (see supplementary S1 Appendix and S1 Dataset) can be used to confirm this behaviour over and beyond the range of observed constituent concentrations. In fact, we would only expect the typical green-peaked reflectance signature of the open Baltic Sea to change when surface accumulations of cyanobacteria significantly enhance near-infrared reflectance.
Combining our observations made thus far, we find that the pelagic zone of the Baltic Sea presents a relatively unique combination of weak particulate backscattering originating primarily from organic particulates, and efficient light absorption by CDOM and water in a large part of the UV to NIR domain. This results in weakly reflecting waters where the absorption signature of phytoplankton is easily masked by CDOM absorption. The implication for remote sensing is that algorithms designed to retrieve phytoplankton Chla biomass by isolating the pigment absorption properties may prove considerably less sensitive in quantifying the pigment than algorithms which are sensitive to light scattering by suspended matter. Because the open Baltic Sea particle population is largely associated with phytoplankton (at least in the productive season), there is a real risk of developing the latter type of algorithm when statistical performance, rather than specificity to phytoplankton pigment in the presence of non-phytoplankton scattering matter, is used as the primary performance criterion.
The dataset described here features rich seasonal diversity in optical properties but also allows us to determine where further investigations of optical properties and associated remote sensing challenges, will be useful. A few observations on where further challenges lie, are given here. In terms of seasonality, our observations focus on the peak of biomass production in spring and early summer, with some additional late summer (august) observations. These periods of peak phytoplankton abundance exhibit significant seasonal differences in IOPs and derived optical-biogeochemical models, which are governed by phytoplankton particles over a wide range of Chla and POC concentrations. It is not yet clear whether non-phytoplankton particulates, in particular those with higher scattering efficiency, play a more significant role in the absorption and scattering budget of the open Baltic Sea when we consider other periods than those of peak periods of phytoplankton growth. In other seasons, including the early summer minimum observed in May-June, we expect to observe further variability in IOPs, either increasing the observed complexity in optical properties for remote sensing algorithm development or, hopefully, leading to a better characterization of gradual optical transitions. Latebloom succession of heterotrophic plankton is also poorly optically characterized at present, and would likely increase the relative importance of a nap (λ) outside peak phytoplankton growth periods. Similarly, autochthonous production of CDOM is associated with bloom degradation and detailed studies of seasonal autochthonous DOM production would lead to better characterization of (regional) relationships between aCDOM and dissolved matter composition than established here. For the benefit of developing seasonally valid optical monitoring strategies it would therefore be highly useful to include IOP measurements in regular sampling activities, even if focussed on relatively few observation stations.
The role of small versus large sized particles, flocculation, and colony formation could not be defined in this study of bulk (non-particle specific) optical properties. The particle size distribution is known to influence pigment absorption efficiency [43] and would likely help explain the variability observed in the pigment packaging effect, both within and between seasons (Fig 12). A weak dependence of b Ã p;TSM ð532Þ and b Ã bp;TSM ð532Þ on the inorganic fraction of TSM (Figs 14 and 17D) may be associated with aging of detrital organic matter in summer stratified surface water, where flocculation associates inorganic matter with larger particles for which we expect lower mass-specific backscattering. Total particulate scattering was higher in summer than in spring whereas backscattering was of similar intensity (Figs 15 and 16). While it is impossible to determine the particle-specific backscatter in our data set because key measurements are lacking (e.g. using flow cytometry, laser diffraction, particle size distributions to obtain individual particle properties), it is clear that summer (cyanobacteria-dominated) particle population were associated with significantly lower backscattering ratios. Cyanobacteria have limited internal structure and are therefore expected to have weak backscatter [56], although species which develop gas vacuoles would elevate backscattering efficiency [57]. There is still controversy in recent literature regarding the backscattering efficiency that should be associated with cyanobacteria blooms, and future remote sensing studies should ideally take into account the morphology of the dominant species before assigning scattering IOPs from the limited body of literature available on this topic.
In conclusion, optical and biogeochemical observations made over a five-year period in productive seasons of the Baltic Sea indicate a clear requirement for bio-optical models and remote sensing methods to adopt optical seasonality. Chla appears the most seasonally variant indicator of the standing stock of phytoplankton in the Baltic Sea. The optical signature of the pigment is severely masked by CDOM, and particulate scattering is too weak to bring out the pigment absorption signature in the apparent optical properties (K d , R rs ) in all but peak bloom conditions. From a purely statistical perspective, TSM and POC are seasonally more robust indicators of the phytoplankton distribution in the pelagic Baltic Sea, compared to Chla. It is interesting that for the Baltic Sea, the use of a single SIOP for phytoplankton using POC as currency may represent the phytoplankton community better across seasons than Chl-a would, and this result may interest e.g. biogeochemical modellers. This does, however, not imply that Chla should be abandoned as an indicator of phytoplankton distributions, particularly since it is most commonly measured in situ. Instead, the sensitivity of optical remote sensing approaches to the seasonally changing role of Chla in the phytoplankton community should be examined, and methods should be developed to optimize separation of absorption by CDOM and particulate matter. Existing Chla retrieval algorithms should also be tested for their sensitivity to the in situ Chla concentration, rather than a correlative response to TSM, as the latter would render them sensitive to overestimation in near-coastal regions and river plumes. The supplementary materials to this paper will support detailed optical-biogeochemical modelling of the Baltic Sea, for which thus far only sparse observations have been made available. Furthermore, the included set of spectral water-leaving reflectance (S1 Appendix and S1 Dataset) supports direct testing of existing and future remote sensing algorithms aiming at retrieval of Chla, TSM, and POC.